The opaque-type-resolution help says "if set before loading symbols",
but I don't think this is accurate. As far as I know, this resolution
can be done at any time.
This patch cleans up the help, also shortening it to less than 80
characters.
Approved-By: Eli Zaretskii <eliz@gnu.org>
This patch ensures that all ordinary help strings are wrapped at 80
columns. For the most part this consists of changing code like this
(note the embedded \n and the trailing backslash without a newline):
-Manage the space-separated list of debuginfod server URLs that GDB will query \
-when missing debuginfo, executables or source files.\nThe default value is \
-copied from the DEBUGINFOD_URLS environment variable."),
... to end each line with \n\, like:
+Manage the space-separated list of debuginfod server URLs that GDB will\n\
+query when missing debuginfo, executables or source files.\n\
+The default value is copied from the DEBUGINFOD_URLS environment variable."),
Approved-By: Eli Zaretskii <eliz@gnu.org>
If you invoke "help function _caller_is", you'll see that the help
text is indented strangely. The fix for this is to add a call to
gdbpy_fix_doc_string_indentation in the appropriate spot, as is
already done for Python commands and parameters.
A customer noted that there is no way to prevent the "current language
does not match this frame" warning. This patch adds a new setting to
allow this warning to be suppressed.
Reviewed-By: Eli Zaretskii <eliz@gnu.org>
Approved-By: Andrew Burgess <aburgess@redhat.com>
Fixed missing operator % on xmethod matcher registration output and, as
suggested on bug 32532, converted both uses of operator % to str.format.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=32352
Change-Id: Ic471516292c2f1d6d1284aaeaea3ec14421decb8
Move note sections after .rodata section so that note sections are
placed in the same PT_LOAD segment together with .rodata section,
instead of a separate PT_LOAD segment.
PR ld/32341
* scripttempl/misc-sections.sc: Move note sections to ...
* scripttempl/elf.sc: Here, after .rodata section.
* testsuite/ld-elf/pr32341.d: New file.
* testsuite/ld-elf/pr32341.s: Likewise.
Co-Authored-By: Nick Clifton <nickc@redhat.com>
Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
I recently came across a case where a compiler would emit a CU with an
empty name. In such case, the attribute object constructed by GDB will
return nullptr when as_string is called. One place is not checking for
this possibility. As a result, loading such binary results in a GDB
crash:
$ gdb -q a.out
Reading symbols from a.out...
Fatal signal: Segmentation fault
----- Backtrace -----
[...]
0x742f4dd8afab __strcmp_avx2
../sysdeps/x86_64/multiarch/strcmp-avx2.S:283
0x58593704a0bc prepare_one_comp_unit
../../gdb/dwarf2/read.c:21842
0x585937053fd9 process_psymtab_comp_unit
../../gdb/dwarf2/read.c:4633
0x585937053fd9 _ZN23cooked_index_debug_info11process_cusEmN9__gnu_cxx17__normal_iteratorIPSt10unique_ptrI18dwarf2_per_cu_data26dwarf2_per_cu_data_deleterESt6vectorIS5_SaIS5_EEEESA_
../../gdb/dwarf2/read.c:4943
[...]
---------------------
A fatal error internal to GDB has been detected, further
debugging is not possible. GDB will now terminate.
This is a bug, please report it. For instructions, see:
<https://www.gnu.org/software/gdb/bugs/>.
Segmentation fault (core dumped)
This seems to be a regression introduced by the following commit:
commit 00105aa1c4
Date: Tue Sep 24 10:24:22 2024 +0200
[gdb/symtab] Don't expand non-Ada CUs for info exceptions
This patch fixes this issue by checking if attr->as_string returns
nullptr.
Change-Id: I78fe7a090f0bd1045b8cb2f8d088a8d6cf57fe1c
Approved-By: Andrew Burgess <aburgess@redhat.com>
Approved-By: Tom Tromey <tom@tromey.com>
Implement extension_language_ops::find_objfile_from_buildid within
GDB's Python API. Doing this allows users to write Python extensions
that can help locate missing objfiles when GDB opens a core file. A
handler might perform some project- or site-specific actions to find a
missing objfile. Or might provide some project- or site-specific
advice to the user on how they can obtain the missing objfile.
The implementation is very similar to the approach taken in:
commit 8f6c452b5a
Date: Sun Oct 15 22:48:42 2023 +0100
gdb: implement missing debug handler hook for Python
The following new commands are added as commands implemented in
Python, this is similar to how the Python missing debug and unwinder
commands are implemented:
info missing-objfile-handlers
enable missing-objfile-handler LOCUS HANDLER
disable missing-objfile-handler LOCUS HANDLER
To make use of this extension hook a user will create missing objfile
handler objects, and registers these handlers with GDB. When GDB
opens a core file and encounters a missing objfile each handler is
called in turn until one is able to help. Here is a minimal handler
that does nothing useful:
import gdb
import gdb.missing_objfile
class MyFirstHandler(gdb.missing_objfile.MissingObjfileHandler):
def __init__(self):
super().__init__("my_first_handler")
def __call__(self, pspace, build_id, filename):
# This handler does nothing useful.
return None
gdb.missing_objfile.register_handler(None, MyFirstHandler())
Returning None from the __call__ method tells GDB that this handler
was unable to find the missing objfile, and GDB should ask any other
registered handlers.
Possible return values from a handler:
- None: This means the handler couldn't help. GDB will call other
registered handlers to see if they can help instead.
- False: The handler has done all it can, but the objfile couldn't
be found. GDB will not call any other handlers, and will
continue without the objfile.
- True: The handler has installed the objfile into a location where
GDB would normally expect to find it. GDB should repeat its
normal lookup process and the objfile should now be found.
- A string: The handler can return a filename, which is the missing
objfile. GDB will load this file.
Handlers can be registered globally, or per program space. GDB checks
the handlers for the current program space first, and then all of the
global handles. The first handler that returns a value that is not
None, has "handled" the missing objfile, at which point GDB continues.
The implementation of this feature is mostly straight forward. I have
reworked some of the missing debug file related code so that it can be
shared with this feature. E.g. gdb/python/lib/gdb/missing_files.py is
mostly content moved from gdb/python/lib/gdb/missing_debug.py, but
updated to be more generic. Now gdb/python/lib/gdb/missing_debug.py
and the new file gdb/python/lib/gdb/missing_objfile.py both call into
the missing_files.py file.
For gdb/python/lib/gdb/command/missing_files.py this is even more
extreme, gdb/python/lib/gdb/command/missing_debug.py is completely
gone now and gdb/python/lib/gdb/command/missing_files.py provides all
of the new commands in a generic way.
I have made one change to the existing Python API, I renamed the
attribute Progspace.missing_debug_handlers to
Progspace.missing_file_handlers. I don't see this as too
problematic. This attribute was only used to implement the missing
debug feature and was never documented beyond the fact that it
existed. There was no reason for users to be touching this attribute.
Reviewed-By: Eli Zaretskii <eliz@gnu.org>
Add a new ext_lang_find_objfile_from_buildid function which is called
from find_objfile_by_build_id and gives extension languages a chance
to find missing objfiles.
This commit adds the ext_lang_find_objfile_from_buildid function and
the extension_language_ops::find_objfile_from_buildid() hook, but does
not implement the hook for any extension languages, that will come in
the next commit.
This commit does rewrite find_objfile_by_build_id (build-id.c) to call
the new hook though. The basic steps of find_objfile_by_build_id are
now this:
1. Try to find the missing objfile using the build-id by looking in
the debug-file-directory's .build-id/ sub-directory. If we find the
file then we're done.
2. Ask debuginfod to download the missing file for us. If we
download the file successfully then we're done.
3. Ask the extension language hook to find the file for us. If the
extension language asks us to try again then we repeat step (1) only
and if we still don't have the file, we move to step (4). If the
extension language told us where the file is then we use that file
and we're done.
4. We didn't find the file. Carry on without it.
Only step (3) is new in this logic, everything else was already done.
There are no tests added here as we can't currently write an extension
language callback. The next commit will add the tests.
Approved-By: Tom Tromey <tom@tromey.com>
In preparation for later commits in this series, rename
ext_lang_missing_debuginfo_result to ext_lang_missing_file_result.
A later commit will add additional Python APIs to handle different
types of missing files beyond just debuginfo.
This is just a rename commit, there should be no functional changes
after this commit.
Approved-By: Tom Tromey <tom@tromey.com>
When opening a core-file GDB is able to use debuginfod to download the
executable that matches the core-file if GDB can find a build-id for
the executable in the core-file.
In this case GDB calls debuginfod_exec_query to download the
executable and GDB prints a message like:
Downloading executable for /path/to/core-file...
which makes sense in that case.
For a long time GDB has also had the ability to download memory-mapped
files and shared libraries when opening a core-file. However, recent
commits have made these cases more likely to trigger, which is a good
thing, but the messaging from GDB in these cases is not ideal. When
downloading a memory-mapped file GDB prints:
Downloading executable for /path/to/memory-mapped-file
And for a shared library:
Downloading executable for /path/to/libfoo.so
These last two messages could, I think, be improved.
I propose making two changes. First, I suggest instead of using
/path/to/core-file in the first case, we use the name of the
executable that GDB is fetching. This makes the messaging consistent
in that we print the name of the file we're fetching rather than the
name of the file we're fetching something for.
I further propose that we replace 'executable for' with the more
generic word 'file'. The messages will then become:
Downloading file /path/to/exec-file...
Downloading file /path/to/memory-mapped-file...
Downloading file /path/to/libfoo.so...
I think these messages are clearer than what we used to have, and they
are consistent in that we name the thing being downloaded in all
cases.
There is one tiny problem. The first case relies on GDB knowing the
name of the executable it wants to download. The only place we can
currently get that from is, I think, the memory-mapped file list.
[ ASIDE: There is `bfd_core_file_failing_command` which reports the
executable and argument list from the core file, but this
information is not ideal for this task. First, the executable and
arguments are merged into a single string, and second, the string is
a relatively short, fixed length string, so the executable name is
often truncated. For these reasons I don't consider fetching the
executable name using this bfd function as a solution. ]
We do have to consider the case that the core file does not have any
mapped file information. This shouldn't ever be the case for a Linux
target, but it's worth considering.
[ ASIDE: I mention Linux specifically because this only becomes a
problem if we try to do a lookup via debuginfod, which requires that
we have build-ids available. Linux has special support for
embedding build-ids into the core file, but I'm not sure if other
kernels do this. ]
For the unlikely edge case of a core-file that has build-ids, but
doesn't have any mapped file information then I propose that we
synthesis a filename like: 'with build-id xxxxxx'. We would then see
a message like:
Downloading file with build-id xxxxxx...
Where 'xxxxxx' would be replaced by the actual build-id.
This isn't ideal, but I think is good enough, and, as I said, I think
this case is not going to be hit very often, or maybe at all.
We already had some tests that emitted two of the above messages,
which I've updated, these cover the mapped-file and shared library
case.
The message about downloading the exec for the core-file is actually
really hard to trigger now as usually the exec will also appear in the
memory-mapped file list and GDB will download the file at this stage.
Then when GDB needs the executable for loading the symbols it'll ask
debuginfod, and debuginfod will find the file in its cache, and so no
message will be printed.
If anyone has any ideas about how to trigger this case then I'm happy
to add additional tests.
Approved-By: Tom Tromey <tom@tromey.com>
This test checks that GDB is able to load DWARF information when
.debug_aranges has a section address size that is set to 0.
This test was originally written by Jan Kratochvil to test commit
927aa2e778 from 2017, titled "DWARF-5: .debug_names index consumer".
This test was originally written using a static .S file and has
been present in the Fedora tree for a long time.
If dwarf2/aranges.c is modified to turn off the address_size check,
GDB will crash with SIGFPE when loading the executable with address
size set to zero.
I modified the DWARF assembler to make it possible to set the address
size to zero in a .debug_aranges section and used the DWARF assembler
to produce the assembly file.
Co-Authored-By: Jan Kratochvil <jan.kratochvil@redhat.com>
Approved-by: Kevin Buettner <kevinb@redhat.com>
This function doesn't seem so useful, use `process_info::pid` directly
instead.
Change-Id: I55d592f38b32a197957ed4c569993cd23a818cb4
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
This function doesn't seem so useful, use `thread_info:🆔:pid`
directly instead.
Change-Id: I7450c4223e5b0bf66788eeb5b070ab6f5287f798
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
This function doesn't seem so useful. Use `thread_info:🆔:lwp`
directly.
Change-Id: Ib4a86eeeee6c1342bc1c092f083589ce28009be1
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
This function doesn't seem so useful. Use `thread_info::id` directly.
Change-Id: I158cd06a752badd30f68424e329aa42d275e43b7
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
This function doesn't seem so useful. Use `thread_info::id` directly.
Change-Id: I4ae4e7baa44e09704631a1c3a5a66e5b8b5a3594
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
I think it just makes things more obscure. Use `thread_info::id`
directly instead.
Change-Id: I141d5fb08ebf45c13cc32c4bba62773249fcb356
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
Remove the `get_thread_process` function, use `thread_info::process`
instead.
In `server.cc`, use `current_process ()` instead of going through the
current thread.
Change-Id: Ifc61d65852e392d154b854a45d45df584ab3922e
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
Same idea as the previous patch, but for `remove_thread`.
Change-Id: I7e227655be5fcf29a3256e8389eb32051f27882d
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
Since thread_info objects are now basically children of process_info
objects, I think that makes sense.
Change-Id: I7f0a67b921b468e512851cb2f36015d1003412d7
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
In a few spots, we need to get to a process from a thread. Having a
backlink from thread to process is cheap and makes the operation
trivial, add it.
Change-Id: I8a94b2919494b1dcaf954de2246386794308aa82
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
Remove this overload, prefer to use `process_info::for_each_thread`. In
many instances, the `process_info` is already available, so this saves a
map lookup. In other instances, add the `process_info` lookup at the
call site.
In `linux-arm-low.cc` and `win32-i386-low.cc`, use `current_process ()`
instead of `current_thread->id.pid ()`. I presume that if
`current_process ()` and `current_thread` don't match, it's a bug
orthogonal to this change.
Change-Id: I751ed497cb1f313cf937b35125151bee9316fc51
Reviewed-By: Tankut Baris Aktemur <tankut.baris.aktemur@intel.com>
LoongArch doesn't implement the hook gdbarch_remove_non_address_bits, so
there is no need to use the hook in gdb/loongarch-linux-nat.c.
Approved-By: Luis Machado <luis.machado@arm.com>
When GDB is unable to read an objfile, it prints the error message "I'm
sorry Dave, I can't do that. Symbol format `%s' unknown.". While it is a
great reference, an end user won't have much information about the
problem.
So far this wasn't much of a problem, as it is very uncommon for GDB to
be unable to read an objfile. However, a future patch will allow users
to selectively disable support to some formats, making it somewhat
expected that the message will be seen by end users.
This commit makes the end message more informative and direct.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=13299
Approved-By: Tom Tromey <tom@tromey.com>
This patch introduces a new operand flag OPD_F_UNSIGNED to signal that
the immediate value should be treated as an unsigned value. The default
signedness of immediate operands is signed.
The current space optmization on enum aarch64_opn_qualifier forced its
encoding using an unsigned char. This "hard-coded" optimization has the
bad consequence of making the array of such enums being completely
unreadable when debugging with GDB because the enum type is lost along
the way.
Keeping this space optimization, and the enum type as well, is possible
when the declaration of the enum is tagged with attribute((packed)).
attribute((packed)) is a GNU extension, and is wrapped in the macro
ATTRIBUTE_PACKED (defined in ansidecl.h), and should be used instead.
The enum aarch64_opnd_qualifiers in include/opcode/aarch64.h needs to
stay in sync with the array of struct operand_qualifier_data which
defines various properties for the different type of operands. For
instance, for:
- registers: the size of the register, the number of elements.
- immediates: lower and upper bits to determine the range of values.
In the recent commit:
commit 31ada87f91
Date: Wed Nov 6 22:18:55 2024 +0000
gdb: fixes and tests for the 'edit' command
the new gdb.base/basic-edit-cmd.exp was added. The Linaro CI
highlighted an issue with the test which I failed to address before
pushing the above commit.
Part of the test loads a file into GDB and then uses the 'edit'
command with no arguments to edit the default location. This default
location is expected to be the 'main' function.
On my local machine the line reported is the opening '{' of main, and
that is what the test checks for.
The Linaro CI though appears to see the first code line of main.
I think either result is fine as far as this test is concerned, so
I've expanded the test regexp to check for either line number. This
should make the CI testing happy again.
This commit was inspired by this mailing list post:
https://inbox.sourceware.org/gdb-patches/osmtfvf5xe3yx4n7oirukidym4cik7lehhy4re5mxpset2qgwt@6qlboxhqiwgm
When reviewing that patch, the first thing I wanted to do was add some
tests for the 'edit' command because, as far as I can tell, there are
no real tests right now.
The approach I've taken for testing is to override the EDITOR
environment variable, setting this to just 'echo'. Now when the
'edit' command is run, instead of entering an interactive editor, the
shell instead echos back the arguments that GDB is trying to pass to
the editor. The output might look like this:
(gdb) edit
+22 /tmp/gdb/testsuite/gdb.base/edit-cmd.c
(gdb)
We can then test this like any other normal command. I then wrote
some basic tests covering a few situations like, using 'edit' before
the inferior is started. Using 'edit' without any arguments, and
using 'edit' with a line number argument.
There are plenty of cases that are still not tested, for example, the
test program only has a single source file for example. But we can
always add more tests later.
I then used these tests to validate the fix proposed in the above
patch.
The patch above does indeed fix some cases, specifically, when GDB
stops at a location (e.g. a breakpoint location) and then the 'edit'
command without any arguments is fixed. But using the 'list' command
to show some other location, and then 'edit' to edit the just listed
location broken before and after the above patch.
I am instead proposing this alternative patch which I think fixes more
cases. When GDB stops at a location then 'edit' with no arguments
should correctly edit the current line. And using 'list XX' to list a
specific location, followed by 'edit' should also now edit the just
listed location.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=17669
Co-Authored-By: Lluís Batlle i Rossell <viric@viric.name>
Approved-By: Tom Tromey <tom@tromey.com>
After commit 2e60790cf7 "Remove the
paramstuff word" there is no caller left of the static find function
in doc/chew.c, so it should be removed.
* doc/chew.c (find): Remove.
For any arm elf target, disable an old piece of code that forced disassembly to
disassemble for 'unknown architecture' which once upon a time meant it would
disassemble ANY arm instruction. This is no longer true with the addition of
Armv8.1-M Mainline, as there are conflicting encodings for different thumb
instructions.
BFD however can detect what architecture the object file was assembled for
using information in the notes section. So if available, we use that,
otherwise we default to the old 'unknown' behaviour.
With the changes above code, a mode changing 'bx lr' assembled for armv4 with
the option --fix-v4bx will result in an object file that is recognized by bfd
as one for the armv4 architecture. The disassembler now disassembles this
encoding as a BX even for Armv4 architectures, but warns the user when
disassembling for Armv4 that this instruction is only valid from Armv4T
onwards.
Remove the unused and wrongfully defined ARM_ARCH_V8A_CRC, and
define and use a ARM_ARCH_V8R_CRC to make sure instructions enabled by
-march=armv8-r+crc are disassembled correctly.
Patch up some of the tests cases, see a brief explanation for each below.
inst.d:
This test checks the assembly & disassembly of basic instructions in armv3m. I
changed the expected behaviour for teqp, cmnp cmpp and testp instructions to
properly print p when disassembling, whereas before, in the 'unknown' case it
would disassemble these as UNPREDICTABLE as they were changed in later
architectures.
nops.d:
Was missing an -march, added one to make sure we were testing the right
behavior of NOP<c> instructions.
unpredictable.d:
Was missing an -march, added armv6 as that reproduced the behaviour being
tested.
We don't seem to support any m-profile assembly/disassembly tests for wince or
pe, so skipping the pacbti one too.
The pr29494 test needs to be skipped because it uses assembly syntax that is
not supported in wince/pe like for instance eabi_attribute directives.
The ARM simulator is no longer able to simulator modern ARM cores, so it
is being deprecated. Once this change has been active for a while - and
assuming that no problems have been found - the ARm simulator codebase
will be removed.
